Proposed Statistical Methods for Signal Detection of Adverse Medical Device Events

Abstract

This paper proposes several statistical approaches for analyzing spontaneous reports of adverse medical device events. The chi-square statistic is used to detect a sudden increase in reports from pairwise comparison of numbers of reports per month. The Cox-Stuart nonparametric trend test is used to detect a gradual, increasing trend in adverse events reports over time. The negative binomial probability model is also used to assess sudden increases by setting threshold values. In this paper, only numerator data (reports of adverse events), not denominator data (device use), are used. The Box-Jenkins ARIMA time series model failed to fit the observed data successfully due to the extremely irregular distributions, including many zeros, of the observed data.